<![CDATA[Breast masses in mammograms are important to segment for computer-aided diagnosis (CAD) to enhance early detection and treatment decisions. Current approaches face challenges in segmenting lesions with low lesion-to-tissue contrast and diverse textures, resulting in misclassification or poor segmentation accuracy. To overcome this challenge, this paper introduces BReMS-Net, a multi-stage segmentation network to improve contextual learning and refined boundaries. We used an MBA-Net backbone with two major components: a Multi-scale Hybrid Dilated Convolution (MHD) module to extract multi-scale contextual features, and a Boundary Feature Auxiliary (BFA) module to strengthen boundary representations via coarse-to-fine feature fusion. Furthermore, a lightweight Prediction-Guided Refinement Module (PRM) uses initial predictions to produce attention maps, remove background clutter, and progressively refine boundary areas. The model has been evaluated on a cross-dataset basis, trained on the CBIS-DDSM dataset and tested on the INbreast dataset, and the results show that the BReMS-Net produces a Dice coefficient of 93.12% and an HD95 of 0.9826, which demonstrate competitive performance compared to several state-of-the-art deep learning methods. These results underline its generalization and robustness. Overall, the framework provides a robust and efficient approach to breast mass segmentation and has important implications for the performance and clinical relevance of automatic breast cancer diagnosis systems.]]>
